Odd width variation of scale down by 2 for subsampling

R=dhrosa@google.com, harryjin@google.com
BUG=libyuv:538

Review URL: https://codereview.chromium.org/1558093003 .

README.chromium

 Name: libyuv
 URL: http://code.google.com/p/libyuv/
-Version: 1562
+Version: 1563
 License: BSD
 License File: LICENSE
 

include/libyuv/scale_row.h

@@ -146,6 +146,8 @@ void ScaleRowDown2Linear_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
                               uint16* dst, int dst_width);
 void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
                         uint8* dst, int dst_width);
+void ScaleRowDown2Box_Odd_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst, int dst_width);
 void ScaleRowDown2Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
                            uint16* dst, int dst_width);
 void ScaleRowDown4_C(const uint8* src_ptr, ptrdiff_t src_stride,
@@ -270,12 +272,16 @@ void ScaleRowDown2Linear_Any_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
                                    uint8* dst_ptr, int dst_width);
 void ScaleRowDown2Box_Any_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
                                 uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Box_Odd_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
+                                uint8* dst_ptr, int dst_width);
 void ScaleRowDown2_Any_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
                             uint8* dst_ptr, int dst_width);
 void ScaleRowDown2Linear_Any_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
                                   uint8* dst_ptr, int dst_width);
 void ScaleRowDown2Box_Any_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
                                uint8* dst_ptr, int dst_width);
+void ScaleRowDown2Box_Odd_AVX2(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst_ptr, int dst_width);
 void ScaleRowDown4_Any_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
                              uint8* dst_ptr, int dst_width);
 void ScaleRowDown4Box_Any_SSSE3(const uint8* src_ptr, ptrdiff_t src_stride,
@@ -431,6 +437,8 @@ void ScaleRowDown2Linear_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                                   uint8* dst, int dst_width);
 void ScaleRowDown2Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                                uint8* dst, int dst_width);
+void ScaleRowDown2Box_Odd_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
+                               uint8* dst, int dst_width);
 void ScaleRowDown4_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,
                             uint8* dst_ptr, int dst_width);
 void ScaleRowDown4Box_Any_NEON(const uint8* src_ptr, ptrdiff_t src_stride,

include/libyuv/version.h

@@ -11,6 +11,6 @@
 #ifndef INCLUDE_LIBYUV_VERSION_H_  // NOLINT
 #define INCLUDE_LIBYUV_VERSION_H_
 
-#define LIBYUV_VERSION 1562
+#define LIBYUV_VERSION 1563
 
 #endif  // INCLUDE_LIBYUV_VERSION_H_  NOLINT

source/planar_functions.cc

@@ -691,29 +691,41 @@ int I420Blend(const uint8* src_y0, int src_stride_y0,
     }
   }
 #endif
+  if (!IS_ALIGNED(width, 2)) {
+    ScaleRowDown2 = ScaleRowDown2Box_Odd_C;
+  }
 #if defined(HAS_SCALEROWDOWN2_NEON)
   if (TestCpuFlag(kCpuHasNEON)) {
+    ScaleRowDown2 = ScaleRowDown2Box_Odd_NEON;
+    if (IS_ALIGNED(width, 2)) {
       ScaleRowDown2 = ScaleRowDown2Box_Any_NEON;
       if (IS_ALIGNED(halfwidth, 16)) {
         ScaleRowDown2 = ScaleRowDown2Box_NEON;
       }
     }
+  }
 #endif
 #if defined(HAS_SCALEROWDOWN2_SSSE3)
   if (TestCpuFlag(kCpuHasSSSE3)) {
+    ScaleRowDown2 = ScaleRowDown2Box_Odd_SSSE3;
+    if (IS_ALIGNED(width, 2)) {
       ScaleRowDown2 = ScaleRowDown2Box_Any_SSSE3;
       if (IS_ALIGNED(halfwidth, 16)) {
         ScaleRowDown2 = ScaleRowDown2Box_SSSE3;
       }
     }
+  }
 #endif
 #if defined(HAS_SCALEROWDOWN2_AVX2)
   if (TestCpuFlag(kCpuHasAVX2)) {
+    ScaleRowDown2 = ScaleRowDown2Box_Odd_AVX2;
+    if (IS_ALIGNED(width, 2)) {
       ScaleRowDown2 = ScaleRowDown2Box_Any_AVX2;
       if (IS_ALIGNED(halfwidth, 32)) {
         ScaleRowDown2 = ScaleRowDown2Box_AVX2;
       }
     }
+  }
 #endif
 
   // Row buffer for intermediate alpha pixels.

source/scale_any.cc

@@ -55,12 +55,29 @@ CANY(ScaleARGBFilterCols_Any_NEON, ScaleARGBFilterCols_NEON,
                      dst_ptr + n * BPP, r);                                    \
     }
 
+// Fixed scale down for odd source width.  Used by I420Blend subsampling.
+// Since dst_width is (width + 1) / 2, this function scales one less pixel
+// and copies the last pixel.
+#define SDODD(NAMEANY, SCALEROWDOWN_SIMD, SCALEROWDOWN_C, FACTOR, BPP, MASK)   \
+    void NAMEANY(const uint8* src_ptr, ptrdiff_t src_stride,                   \
+                 uint8* dst_ptr, int dst_width) {                              \
+      int r = (int)((unsigned int)(dst_width - 1) % (MASK + 1));               \
+      int n = dst_width - r;                                                   \
+      if (n > 0) {                                                             \
+        SCALEROWDOWN_SIMD(src_ptr, src_stride, dst_ptr, n);                    \
+      }                                                                        \
+      SCALEROWDOWN_C(src_ptr + (n * FACTOR) * BPP, src_stride,                 \
+                     dst_ptr + n * BPP, r);                                    \
+    }
+
 #ifdef HAS_SCALEROWDOWN2_SSSE3
 SDANY(ScaleRowDown2_Any_SSSE3, ScaleRowDown2_SSSE3, ScaleRowDown2_C, 2, 1, 15)
 SDANY(ScaleRowDown2Linear_Any_SSSE3, ScaleRowDown2Linear_SSSE3,
       ScaleRowDown2Linear_C, 2, 1, 15)
 SDANY(ScaleRowDown2Box_Any_SSSE3, ScaleRowDown2Box_SSSE3, ScaleRowDown2Box_C,
       2, 1, 15)
+SDODD(ScaleRowDown2Box_Odd_SSSE3, ScaleRowDown2Box_SSSE3,
+      ScaleRowDown2Box_Odd_C, 2, 1, 15)
 #endif
 #ifdef HAS_SCALEROWDOWN2_AVX2
 SDANY(ScaleRowDown2_Any_AVX2, ScaleRowDown2_AVX2, ScaleRowDown2_C, 2, 1, 31)
@@ -68,6 +85,8 @@ SDANY(ScaleRowDown2Linear_Any_AVX2, ScaleRowDown2Linear_AVX2,
       ScaleRowDown2Linear_C, 2, 1, 31)
 SDANY(ScaleRowDown2Box_Any_AVX2, ScaleRowDown2Box_AVX2, ScaleRowDown2Box_C,
       2, 1, 31)
+SDODD(ScaleRowDown2Box_Odd_AVX2, ScaleRowDown2Box_AVX2, ScaleRowDown2Box_Odd_C,
+      2, 1, 31)
 #endif
 #ifdef HAS_SCALEROWDOWN2_NEON
 SDANY(ScaleRowDown2_Any_NEON, ScaleRowDown2_NEON, ScaleRowDown2_C, 2, 1, 15)
@@ -75,6 +94,8 @@ SDANY(ScaleRowDown2Linear_Any_NEON, ScaleRowDown2Linear_NEON,
       ScaleRowDown2Linear_C, 2, 1, 15)
 SDANY(ScaleRowDown2Box_Any_NEON, ScaleRowDown2Box_NEON,
       ScaleRowDown2Box_C, 2, 1, 15)
+SDODD(ScaleRowDown2Box_Odd_NEON, ScaleRowDown2Box_NEON,
+      ScaleRowDown2Box_Odd_C, 2, 1, 15)
 #endif
 #ifdef HAS_SCALEROWDOWN4_SSSE3
 SDANY(ScaleRowDown4_Any_SSSE3, ScaleRowDown4_SSSE3, ScaleRowDown4_C, 4, 1, 7)

source/scale_common.cc

@@ -103,6 +103,28 @@ void ScaleRowDown2Box_C(const uint8* src_ptr, ptrdiff_t src_stride,
   }
 }
 
+void ScaleRowDown2Box_Odd_C(const uint8* src_ptr, ptrdiff_t src_stride,
+                            uint8* dst, int dst_width) {
+  const uint8* s = src_ptr;
+  const uint8* t = src_ptr + src_stride;
+  int x;
+  dst_width -= 1;
+  for (x = 0; x < dst_width - 1; x += 2) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+    dst[1] = (s[2] + s[3] + t[2] + t[3] + 2) >> 2;
+    dst += 2;
+    s += 4;
+    t += 4;
+  }
+  if (dst_width & 1) {
+    dst[0] = (s[0] + s[1] + t[0] + t[1] + 2) >> 2;
+    dst += 1;
+    s += 2;
+    t += 2;
+  }
+  dst[0] = (s[0] + t[0] + 1) >> 1;
+}
+
 void ScaleRowDown2Box_16_C(const uint16* src_ptr, ptrdiff_t src_stride,
                            uint16* dst, int dst_width) {
   const uint16* s = src_ptr;